transform: failed to synchronize: cudaErrorAssert: device-side assert triggered

This problem usually occurs when the loss function is loss.backward() This sentence ：
The first is classification ：
   First, check whether the loss function of multi classification problem is used as the loss function of binary classification , Secondly, check whether the number of categories of the loss function of multiple classifications corresponds , Then if it is a binary loss function binary_cross_entropy（） Remember to add the activation function sigmoid, There are many categories. They come with softmax Never mind the function
Then there is the problem of regression ：
   For the return question , The commonly used loss function is MSEloss, The problems in the title generally do not appear in MSEloss In itself , That is to say, it is mostly the problem of the model . Therefore, this time should be used CPU Run the model to get the real error information , And if used GPU Running it will only cover up the real error information .